Paul A. Clayton (paaronclayton.delete@this.gmail.com) on January 9, 2020 2:15 pm wrote:
> Anon (anon.delete@this.anon.com) on January 8, 2020 8:37 pm wrote:
> [snip]
> > To name one bad feature I will say segments, of all methods of addressing
> > data beyond the native int size Intel probably choose the worst of them.
>
> HP PA-RISC segments seemed okay (though I think using the least significant bits rather
> than the most significant bits to provide an "implicit" segment number might have been
> better). However, this is a theoretical perspective not from experience.
>
> x86 segments mixed address extension and protection and were relatively complex. Having a code




Exactly like you said, x85 segments were used for two different purposes, address extension and memory protection.

Segments used for address extension are definitely very bad, there is no doubt about that. I remember with horror the ancient times when I had to juggle with the large, medium, compact and small memory models.


Segments used only for memory protection should be OK, because they are transparent for software, the same like paging.


In fact I believe that most extra features provided by paging besides memory protection (i.e. dynamic relocation of fragmented memory and virtual memory) are not really needed. If one would not use those features, then segments would be a more efficient way to implement memory protection.


For virtual memory, I am quite sure that it is useless, because during the last 20 years I have always disabled the use of swap on all Linux and FreeBSD servers and workstations, both at home and at work, and this was always an improvement. (But I have always provided enough real memory.)


For the dynamic relocation of fragmented memory, I cannot prove my belief yet, i.e. that alternative approaches are not worse than using paging for relocation, so that the virtual to real address translation could be omitted or simplified, because for that I would have for example to complete an OS kernel able to run any Linux applications while using only identity mappings in the page tables and then show that there are no disadvantages compared to a traditional kernel.


Nevertheless, I am actually working at that.